mDeBERTa-v3-base-mnli-xnli-tl-messages

Model Description

This is a fine-tuned model for text classification, specifically designed for multi-label text categorization across 17 distinct content categories of Telegram messages.

Intended Use & Limitations

This model is designed for classifying text into one or more of the following 17 content categories:

ID	Label	Description
0	admission	University admissions, enrollment procedures
1	achievement	Accomplishments, awards, honors
2	social	Social events, community activities
3	digest	News summaries, digests
4	science	Scientific research, discoveries
5	career	Job opportunities, career development
6	patriotism	National pride, patriotic events
7	sport	Sports events, athletic activities
8	volunteering	Volunteer work, community service
9	competition	Contests, competitions, tournaments, hackathons
10	education	Educational events
11	announcement	Official announcements, notices
12	international_relations	Exchange programs, international cooperation
13	rating	Reviews, rankings, ratings
14	scholarship	Financial aid, scholarships, grants
15	weather_forecast	Weather predictions, forecasts
16	cybersecurity	Security threats

Training Details

Hyperparameters

• Training Epochs: 18
• Batch Size: 8
• Learning Rate: 2e-5 with linear decay
• Max Sequence Length: 512 tokens
• Optimizer: AdamW with weight decay
• Warmup Steps: Integrated in learning rate schedule

Performance

• Best F1 Score: 0.8174 (achieved at step 1248)
• Best Accuracy: 82.05% (achieved at step 1248)

Limitations

1. Domain Specificity: Trained for a classification of a specific Telegram channel's messages
2. Language: Primarily for Russian Telegram text messages
3. Max Length: Limited to 512 tokens